Unverified50% confidenceOpinionExact time
The way safety filters are enforced has side effects: when the refusal curve is gentle, users learn that "just trying again works," which runs counter to the goals of safety design.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedDo not exclude blocked trajectories from training, because filtering out blocked trajectories is still selecting trajectories that obtain high reward by evading the monitor, implicitly inversely selecting blocked trajectories, and due to the nature of softmax, increasing the probability of evasion necessarily comes at the cost of lowering the probability of non-evasion.73% similarUnverifiedUsing code hooks for hard interception is more reliable than placing safety rules in prompts and relying on the model to self-enforce them, because the model can be bypassed or produce hallucinations.71% similarUnverified越狱技术通过特定提示工程手段绕过模型内置安全护栏，使其输出本应被拒绝的内容71% similarUnverifiedUncensored models remove or weaken overly conservative refusal behavior, thereby providing more practical technical output in security scenarios.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1044149API
curl https://kongchang.com/api/v1/knowledge/claims/1044149MCP
get_claim(id=1044149)